Mercurial > jdk8-mips64-public > hotspot / file revision

 /*

  * Copyright (c) 1998, 2014, Oracle and/or its affiliates. All rights reserved.

  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.

  *

  * This code is free software; you can redistribute it and/or modify it

  * under the terms of the GNU General Public License version 2 only, as

  * published by the Free Software Foundation.

  *

  * This code is distributed in the hope that it will be useful, but WITHOUT

  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or

  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License

  * version 2 for more details (a copy is included in the LICENSE file that

  * accompanied this code).

  *

  * You should have received a copy of the GNU General Public License version

  * 2 along with this work; if not, write to the Free Software Foundation,

  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.

  *

  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA

  * or visit www.oracle.com if you need additional information or have any

  * questions.

  *

  */

 #include "precompiled.hpp"

 #include "ci/ciMethodData.hpp"

 #include "compiler/compileLog.hpp"

 #include "libadt/vectset.hpp"

 #include "memory/allocation.inline.hpp"

 #include "opto/addnode.hpp"

 #include "opto/callnode.hpp"

 #include "opto/connode.hpp"

 #include "opto/divnode.hpp"

 #include "opto/idealGraphPrinter.hpp"

 #include "opto/loopnode.hpp"

 #include "opto/mulnode.hpp"

 #include "opto/rootnode.hpp"

 #include "opto/superword.hpp"

 //=============================================================================

 //------------------------------is_loop_iv-------------------------------------

 // Determine if a node is Counted loop induction variable.

 // The method is declared in node.hpp.

 const Node* Node::is_loop_iv() const {

   if (this->is_Phi() && !this->as_Phi()->is_copy() &&

       this->as_Phi()->region()->is_CountedLoop() &&

       this->as_Phi()->region()->as_CountedLoop()->phi() == this) {

     return this;

   } else {

     return NULL;

   }

 }

 //=============================================================================

 //------------------------------dump_spec--------------------------------------

 // Dump special per-node info

 #ifndef PRODUCT

 void LoopNode::dump_spec(outputStream *st) const {

   if (is_inner_loop()) st->print( "inner " );

   if (is_partial_peel_loop()) st->print( "partial_peel " );

   if (partial_peel_has_failed()) st->print( "partial_peel_failed " );

 }

 #endif

 //------------------------------is_valid_counted_loop-------------------------

 bool LoopNode::is_valid_counted_loop() const {

   if (is_CountedLoop()) {

     CountedLoopNode*    l  = as_CountedLoop();

     CountedLoopEndNode* le = l->loopexit();

     if (le != NULL &&

         le->proj_out(1 /* true */) == l->in(LoopNode::LoopBackControl)) {

       Node* phi  = l->phi();

       Node* exit = le->proj_out(0 /* false */);

       if (exit != NULL && exit->Opcode() == Op_IfFalse &&

           phi != NULL && phi->is_Phi() &&

           phi->in(LoopNode::LoopBackControl) == l->incr() &&

           le->loopnode() == l && le->stride_is_con()) {

         return true;

       }

     }

   }

   return false;

 }

 //------------------------------get_early_ctrl---------------------------------

 // Compute earliest legal control

 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {

   assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );

   uint i;

   Node *early;

   if (n->in(0) && !n->is_expensive()) {

     early = n->in(0);

     if (!early->is_CFG()) // Might be a non-CFG multi-def

       early = get_ctrl(early);        // So treat input as a straight data input

     i = 1;

   } else {

     early = get_ctrl(n->in(1));

     i = 2;

   }

   uint e_d = dom_depth(early);

   assert( early, "" );

   for (; i < n->req(); i++) {

     Node *cin = get_ctrl(n->in(i));

     assert( cin, "" );

     // Keep deepest dominator depth

     uint c_d = dom_depth(cin);

     if (c_d > e_d) {           // Deeper guy?

       early = cin;              // Keep deepest found so far

       e_d = c_d;

     } else if (c_d == e_d &&    // Same depth?

                early != cin) { // If not equal, must use slower algorithm

       // If same depth but not equal, one _must_ dominate the other

       // and we want the deeper (i.e., dominated) guy.

       Node *n1 = early;

       Node *n2 = cin;

       while (1) {

         n1 = idom(n1);          // Walk up until break cycle

         n2 = idom(n2);

         if (n1 == cin ||        // Walked early up to cin

             dom_depth(n2) < c_d)

           break;                // early is deeper; keep him

         if (n2 == early ||      // Walked cin up to early

             dom_depth(n1) < c_d) {

           early = cin;          // cin is deeper; keep him

           break;

         }

       }

       e_d = dom_depth(early);   // Reset depth register cache

     }

   }

   // Return earliest legal location

   assert(early == find_non_split_ctrl(early), "unexpected early control");

   if (n->is_expensive()) {

     assert(n->in(0), "should have control input");

     early = get_early_ctrl_for_expensive(n, early);

   }

   return early;

 }

 //------------------------------get_early_ctrl_for_expensive---------------------------------

 // Move node up the dominator tree as high as legal while still beneficial

 Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) {

   assert(n->in(0) && n->is_expensive(), "expensive node with control input here");

   assert(OptimizeExpensiveOps, "optimization off?");

   Node* ctl = n->in(0);

   assert(ctl->is_CFG(), "expensive input 0 must be cfg");

   uint min_dom_depth = dom_depth(earliest);

 #ifdef ASSERT

   if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) {

     dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl);

     assert(false, "Bad graph detected in get_early_ctrl_for_expensive");

   }

 #endif

   if (dom_depth(ctl) < min_dom_depth) {

     return earliest;

   }

   while (1) {

     Node *next = ctl;

     // Moving the node out of a loop on the projection of a If

     // confuses loop predication. So once we hit a Loop in a If branch

     // that doesn't branch to an UNC, we stop. The code that process

     // expensive nodes will notice the loop and skip over it to try to

     // move the node further up.

     if (ctl->is_CountedLoop() && ctl->in(1) != NULL && ctl->in(1)->in(0) != NULL && ctl->in(1)->in(0)->is_If()) {

       if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {

         break;

       }

       next = idom(ctl->in(1)->in(0));

     } else if (ctl->is_Proj()) {

       // We only move it up along a projection if the projection is

       // the single control projection for its parent: same code path,

       // if it's a If with UNC or fallthrough of a call.

       Node* parent_ctl = ctl->in(0);

       if (parent_ctl == NULL) {

         break;

       } else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != NULL) {

         next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control();

       } else if (parent_ctl->is_If()) {

         if (!ctl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {

           break;

         }

         assert(idom(ctl) == parent_ctl, "strange");

         next = idom(parent_ctl);

       } else if (ctl->is_CatchProj()) {

         if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) {

           break;

         }

         assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph");

         next = parent_ctl->in(0)->in(0)->in(0);

       } else {

         // Check if parent control has a single projection (this

         // control is the only possible successor of the parent

         // control). If so, we can try to move the node above the

         // parent control.

         int nb_ctl_proj = 0;

         for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) {

           Node *p = parent_ctl->fast_out(i);

           if (p->is_Proj() && p->is_CFG()) {

             nb_ctl_proj++;

             if (nb_ctl_proj > 1) {

               break;

             }

           }

         }

         if (nb_ctl_proj > 1) {

           break;

         }

         assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call(), "unexpected node");

         assert(idom(ctl) == parent_ctl, "strange");

         next = idom(parent_ctl);

       }

     } else {

       next = idom(ctl);

     }

     if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) {

       break;

     }

     ctl = next;

   }

   if (ctl != n->in(0)) {

     _igvn.hash_delete(n);

     n->set_req(0, ctl);

     _igvn.hash_insert(n);

   }

   return ctl;

 }

 //------------------------------set_early_ctrl---------------------------------

 // Set earliest legal control

 void PhaseIdealLoop::set_early_ctrl( Node *n ) {

   Node *early = get_early_ctrl(n);

   // Record earliest legal location

   set_ctrl(n, early);

 }

 //------------------------------set_subtree_ctrl-------------------------------

 // set missing _ctrl entries on new nodes

 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) {

   // Already set?  Get out.

   if( _nodes[n->_idx] ) return;

   // Recursively set _nodes array to indicate where the Node goes

   uint i;

   for( i = 0; i < n->req(); ++i ) {

     Node *m = n->in(i);

     if( m && m != C->root() )

       set_subtree_ctrl( m );

   }

   // Fixup self

   set_early_ctrl( n );

 }

 //------------------------------is_counted_loop--------------------------------

 bool PhaseIdealLoop::is_counted_loop( Node *x, IdealLoopTree *loop ) {

   PhaseGVN *gvn = &_igvn;

   // Counted loop head must be a good RegionNode with only 3 not NULL

   // control input edges: Self, Entry, LoopBack.

   if (x->in(LoopNode::Self) == NULL || x->req() != 3 || loop->_irreducible) {

     return false;

   }

   Node *init_control = x->in(LoopNode::EntryControl);

   Node *back_control = x->in(LoopNode::LoopBackControl);

   if (init_control == NULL || back_control == NULL)    // Partially dead

     return false;

   // Must also check for TOP when looking for a dead loop

   if (init_control->is_top() || back_control->is_top())

     return false;

   // Allow funny placement of Safepoint

   if (back_control->Opcode() == Op_SafePoint)

     back_control = back_control->in(TypeFunc::Control);

   // Controlling test for loop

   Node *iftrue = back_control;

   uint iftrue_op = iftrue->Opcode();

   if (iftrue_op != Op_IfTrue &&

       iftrue_op != Op_IfFalse)

     // I have a weird back-control.  Probably the loop-exit test is in

     // the middle of the loop and I am looking at some trailing control-flow

     // merge point.  To fix this I would have to partially peel the loop.

     return false; // Obscure back-control

   // Get boolean guarding loop-back test

   Node *iff = iftrue->in(0);

   if (get_loop(iff) != loop || !iff->in(1)->is_Bool())

     return false;

   BoolNode *test = iff->in(1)->as_Bool();

   BoolTest::mask bt = test->_test._test;

   float cl_prob = iff->as_If()->_prob;

   if (iftrue_op == Op_IfFalse) {

     bt = BoolTest(bt).negate();

     cl_prob = 1.0 - cl_prob;

   }

   // Get backedge compare

   Node *cmp = test->in(1);

   int cmp_op = cmp->Opcode();

   if (cmp_op != Op_CmpI)

     return false;                // Avoid pointer & float compares

   // Find the trip-counter increment & limit.  Limit must be loop invariant.

   Node *incr  = cmp->in(1);

   Node *limit = cmp->in(2);

   // ---------

   // need 'loop()' test to tell if limit is loop invariant

   // ---------

   if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?

     Node *tmp = incr;            // Then reverse order into the CmpI

     incr = limit;

     limit = tmp;

     bt = BoolTest(bt).commute(); // And commute the exit test

   }

   if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant

     return false;

   if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant

     return false;

   Node* phi_incr = NULL;

   // Trip-counter increment must be commutative & associative.

   if (incr->is_Phi()) {

     if (incr->as_Phi()->region() != x || incr->req() != 3)

       return false; // Not simple trip counter expression

     phi_incr = incr;

     incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi

     if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant

       return false;

   }

   Node* trunc1 = NULL;

   Node* trunc2 = NULL;

   const TypeInt* iv_trunc_t = NULL;

   if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) {

     return false; // Funny increment opcode

   }

   assert(incr->Opcode() == Op_AddI, "wrong increment code");

   // Get merge point

   Node *xphi = incr->in(1);

   Node *stride = incr->in(2);

   if (!stride->is_Con()) {     // Oops, swap these

     if (!xphi->is_Con())       // Is the other guy a constant?

       return false;             // Nope, unknown stride, bail out

     Node *tmp = xphi;           // 'incr' is commutative, so ok to swap

     xphi = stride;

     stride = tmp;

   }

   // Stride must be constant

   int stride_con = stride->get_int();

   if (stride_con == 0)

     return false; // missed some peephole opt

   if (!xphi->is_Phi())

     return false; // Too much math on the trip counter

   if (phi_incr != NULL && phi_incr != xphi)

     return false;

   PhiNode *phi = xphi->as_Phi();

   // Phi must be of loop header; backedge must wrap to increment

   if (phi->region() != x)

     return false;

   if (trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr ||

       trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1) {

     return false;

   }

   Node *init_trip = phi->in(LoopNode::EntryControl);

   // If iv trunc type is smaller than int, check for possible wrap.

   if (!TypeInt::INT->higher_equal(iv_trunc_t)) {

     assert(trunc1 != NULL, "must have found some truncation");

     // Get a better type for the phi (filtered thru if's)

     const TypeInt* phi_ft = filtered_type(phi);

     // Can iv take on a value that will wrap?

     //

     // Ensure iv's limit is not within "stride" of the wrap value.

     //

     // Example for "short" type

     //    Truncation ensures value is in the range -32768..32767 (iv_trunc_t)

     //    If the stride is +10, then the last value of the induction

     //    variable before the increment (phi_ft->_hi) must be

     //    <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to

     //    ensure no truncation occurs after the increment.

     if (stride_con > 0) {

       if (iv_trunc_t->_hi - phi_ft->_hi < stride_con ||

           iv_trunc_t->_lo > phi_ft->_lo) {

         return false;  // truncation may occur

       }

     } else if (stride_con < 0) {

       if (iv_trunc_t->_lo - phi_ft->_lo > stride_con ||

           iv_trunc_t->_hi < phi_ft->_hi) {

         return false;  // truncation may occur

       }

     }

     // No possibility of wrap so truncation can be discarded

     // Promote iv type to Int

   } else {

     assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");

   }

   // If the condition is inverted and we will be rolling

   // through MININT to MAXINT, then bail out.

   if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!

       // Odd stride

       bt == BoolTest::ne && stride_con != 1 && stride_con != -1 ||

       // Count down loop rolls through MAXINT

       (bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0 ||

       // Count up loop rolls through MININT

       (bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0) {

     return false; // Bail out

   }

   const TypeInt* init_t = gvn->type(init_trip)->is_int();

   const TypeInt* limit_t = gvn->type(limit)->is_int();

   if (stride_con > 0) {

     jlong init_p = (jlong)init_t->_lo + stride_con;

     if (init_p > (jlong)max_jint || init_p > (jlong)limit_t->_hi)

       return false; // cyclic loop or this loop trips only once

   } else {

     jlong init_p = (jlong)init_t->_hi + stride_con;

     if (init_p < (jlong)min_jint || init_p < (jlong)limit_t->_lo)

       return false; // cyclic loop or this loop trips only once

   }

   if (phi_incr != NULL) {

     // check if there is a possiblity of IV overflowing after the first increment

     if (stride_con > 0) {

       if (init_t->_hi > max_jint - stride_con) {

         return false;

       }

     } else {

       if (init_t->_lo < min_jint - stride_con) {

         return false;

       }

     }

   }

   // =================================================

   // ---- SUCCESS!   Found A Trip-Counted Loop!  -----

   //

   assert(x->Opcode() == Op_Loop, "regular loops only");

   C->print_method(PHASE_BEFORE_CLOOPS, 3);

   Node *hook = new (C) Node(6);

   if (LoopLimitCheck) {

   // ===================================================

   // Generate loop limit check to avoid integer overflow

   // in cases like next (cyclic loops):

   //

   // for (i=0; i <= max_jint; i++) {}

   // for (i=0; i <  max_jint; i+=2) {}

   //

   //

   // Limit check predicate depends on the loop test:

   //

   // for(;i != limit; i++)       --> limit <= (max_jint)

   // for(;i <  limit; i+=stride) --> limit <= (max_jint - stride + 1)

   // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride    )

   //

   // Check if limit is excluded to do more precise int overflow check.

   bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge);

   int stride_m  = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1));

   // If compare points directly to the phi we need to adjust

   // the compare so that it points to the incr. Limit have

   // to be adjusted to keep trip count the same and the

   // adjusted limit should be checked for int overflow.

   if (phi_incr != NULL) {

     stride_m  += stride_con;

   }

   if (limit->is_Con()) {

     int limit_con = limit->get_int();

     if ((stride_con > 0 && limit_con > (max_jint - stride_m)) ||

         (stride_con < 0 && limit_con < (min_jint - stride_m))) {

       // Bailout: it could be integer overflow.

       return false;

     }

   } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) ||

              (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) {

       // Limit's type may satisfy the condition, for example,

       // when it is an array length.

   } else {

     // Generate loop's limit check.

     // Loop limit check predicate should be near the loop.

     ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);

     if (!limit_check_proj) {

       // The limit check predicate is not generated if this method trapped here before.

 #ifdef ASSERT

       if (TraceLoopLimitCheck) {

         tty->print("missing loop limit check:");

         loop->dump_head();

         x->dump(1);

       }

 #endif

       return false;

     }

     IfNode* check_iff = limit_check_proj->in(0)->as_If();

     Node* cmp_limit;

     Node* bol;

     if (stride_con > 0) {

       cmp_limit = new (C) CmpINode(limit, _igvn.intcon(max_jint - stride_m));

       bol = new (C) BoolNode(cmp_limit, BoolTest::le);

     } else {

       cmp_limit = new (C) CmpINode(limit, _igvn.intcon(min_jint - stride_m));

       bol = new (C) BoolNode(cmp_limit, BoolTest::ge);

     }

     cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);

     bol = _igvn.register_new_node_with_optimizer(bol);

     set_subtree_ctrl(bol);

     // Replace condition in original predicate but preserve Opaque node

     // so that previous predicates could be found.

     assert(check_iff->in(1)->Opcode() == Op_Conv2B &&

            check_iff->in(1)->in(1)->Opcode() == Op_Opaque1, "");

     Node* opq = check_iff->in(1)->in(1);

     _igvn.hash_delete(opq);

     opq->set_req(1, bol);

     // Update ctrl.

     set_ctrl(opq, check_iff->in(0));

     set_ctrl(check_iff->in(1), check_iff->in(0));

 #ifndef PRODUCT

     // report that the loop predication has been actually performed

     // for this loop

     if (TraceLoopLimitCheck) {

       tty->print_cr("Counted Loop Limit Check generated:");

       debug_only( bol->dump(2); )

     }

 #endif

   }

   if (phi_incr != NULL) {

     // If compare points directly to the phi we need to adjust

     // the compare so that it points to the incr. Limit have

     // to be adjusted to keep trip count the same and we

     // should avoid int overflow.

     //

     //   i = init; do {} while(i++ < limit);

     // is converted to

     //   i = init; do {} while(++i < limit+1);

     //

     limit = gvn->transform(new (C) AddINode(limit, stride));

   }

   // Now we need to canonicalize loop condition.

   if (bt == BoolTest::ne) {

     assert(stride_con == 1 || stride_con == -1, "simple increment only");

     // 'ne' can be replaced with 'lt' only when init < limit.

     if (stride_con > 0 && init_t->_hi < limit_t->_lo)

       bt = BoolTest::lt;

     // 'ne' can be replaced with 'gt' only when init > limit.

     if (stride_con < 0 && init_t->_lo > limit_t->_hi)

       bt = BoolTest::gt;

   }

   if (incl_limit) {

     // The limit check guaranties that 'limit <= (max_jint - stride)' so

     // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.

     //

     Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1);

     limit = gvn->transform(new (C) AddINode(limit, one));

     if (bt == BoolTest::le)

       bt = BoolTest::lt;

     else if (bt == BoolTest::ge)

       bt = BoolTest::gt;

     else

       ShouldNotReachHere();

   }

   set_subtree_ctrl( limit );

   } else { // LoopLimitCheck

   // If compare points to incr, we are ok.  Otherwise the compare

   // can directly point to the phi; in this case adjust the compare so that

   // it points to the incr by adjusting the limit.

   if (cmp->in(1) == phi || cmp->in(2) == phi)

     limit = gvn->transform(new (C) AddINode(limit,stride));

   // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride.

   // Final value for iterator should be: trip_count * stride + init_trip.

   Node *one_p = gvn->intcon( 1);

   Node *one_m = gvn->intcon(-1);

   Node *trip_count = NULL;

   switch( bt ) {

   case BoolTest::eq:

     ShouldNotReachHere();

   case BoolTest::ne:            // Ahh, the case we desire

     if (stride_con == 1)

       trip_count = gvn->transform(new (C) SubINode(limit,init_trip));

     else if (stride_con == -1)

       trip_count = gvn->transform(new (C) SubINode(init_trip,limit));

     else

       ShouldNotReachHere();

     set_subtree_ctrl(trip_count);

     //_loop.map(trip_count->_idx,loop(limit));

     break;

   case BoolTest::le:            // Maybe convert to '<' case

     limit = gvn->transform(new (C) AddINode(limit,one_p));

     set_subtree_ctrl( limit );

     hook->init_req(4, limit);

     bt = BoolTest::lt;

     // Make the new limit be in the same loop nest as the old limit

     //_loop.map(limit->_idx,limit_loop);

     // Fall into next case

   case BoolTest::lt: {          // Maybe convert to '!=' case

     if (stride_con < 0) // Count down loop rolls through MAXINT

       ShouldNotReachHere();

     Node *range = gvn->transform(new (C) SubINode(limit,init_trip));

     set_subtree_ctrl( range );

     hook->init_req(0, range);

     Node *bias  = gvn->transform(new (C) AddINode(range,stride));

     set_subtree_ctrl( bias );

     hook->init_req(1, bias);

     Node *bias1 = gvn->transform(new (C) AddINode(bias,one_m));

     set_subtree_ctrl( bias1 );

     hook->init_req(2, bias1);

     trip_count  = gvn->transform(new (C) DivINode(0,bias1,stride));

     set_subtree_ctrl( trip_count );

     hook->init_req(3, trip_count);

     break;

   }

   case BoolTest::ge:            // Maybe convert to '>' case

     limit = gvn->transform(new (C) AddINode(limit,one_m));

     set_subtree_ctrl( limit );

     hook->init_req(4 ,limit);

     bt = BoolTest::gt;

     // Make the new limit be in the same loop nest as the old limit

     //_loop.map(limit->_idx,limit_loop);

     // Fall into next case

   case BoolTest::gt: {          // Maybe convert to '!=' case

     if (stride_con > 0) // count up loop rolls through MININT

       ShouldNotReachHere();

     Node *range = gvn->transform(new (C) SubINode(limit,init_trip));

     set_subtree_ctrl( range );

     hook->init_req(0, range);

     Node *bias  = gvn->transform(new (C) AddINode(range,stride));

     set_subtree_ctrl( bias );

     hook->init_req(1, bias);

     Node *bias1 = gvn->transform(new (C) AddINode(bias,one_p));

     set_subtree_ctrl( bias1 );

     hook->init_req(2, bias1);

     trip_count  = gvn->transform(new (C) DivINode(0,bias1,stride));

     set_subtree_ctrl( trip_count );

     hook->init_req(3, trip_count);

     break;

   }

   } // switch( bt )

   Node *span = gvn->transform(new (C) MulINode(trip_count,stride));

   set_subtree_ctrl( span );

   hook->init_req(5, span);

   limit = gvn->transform(new (C) AddINode(span,init_trip));

   set_subtree_ctrl( limit );

   } // LoopLimitCheck

   // Check for SafePoint on backedge and remove

   Node *sfpt = x->in(LoopNode::LoopBackControl);

   if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {

     lazy_replace( sfpt, iftrue );

     if (loop->_safepts != NULL) {

       loop->_safepts->yank(sfpt);

     }

     loop->_tail = iftrue;

   }

   // Build a canonical trip test.

   // Clone code, as old values may be in use.

   incr = incr->clone();

   incr->set_req(1,phi);

   incr->set_req(2,stride);

   incr = _igvn.register_new_node_with_optimizer(incr);

   set_early_ctrl( incr );

   _igvn.hash_delete(phi);

   phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );

   // If phi type is more restrictive than Int, raise to

   // Int to prevent (almost) infinite recursion in igvn

   // which can only handle integer types for constants or minint..maxint.

   if (!TypeInt::INT->higher_equal(phi->bottom_type())) {

     Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT);

     nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));

     nphi = _igvn.register_new_node_with_optimizer(nphi);

     set_ctrl(nphi, get_ctrl(phi));

     _igvn.replace_node(phi, nphi);

     phi = nphi->as_Phi();

   }

   cmp = cmp->clone();

   cmp->set_req(1,incr);

   cmp->set_req(2,limit);

   cmp = _igvn.register_new_node_with_optimizer(cmp);

   set_ctrl(cmp, iff->in(0));

   test = test->clone()->as_Bool();

   (*(BoolTest*)&test->_test)._test = bt;

   test->set_req(1,cmp);

   _igvn.register_new_node_with_optimizer(test);

   set_ctrl(test, iff->in(0));

   // Replace the old IfNode with a new LoopEndNode

   Node *lex = _igvn.register_new_node_with_optimizer(new (C) CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));

   IfNode *le = lex->as_If();

   uint dd = dom_depth(iff);

   set_idom(le, le->in(0), dd); // Update dominance for loop exit

   set_loop(le, loop);

   // Get the loop-exit control

   Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));

   // Need to swap loop-exit and loop-back control?

   if (iftrue_op == Op_IfFalse) {

     Node *ift2=_igvn.register_new_node_with_optimizer(new (C) IfTrueNode (le));

     Node *iff2=_igvn.register_new_node_with_optimizer(new (C) IfFalseNode(le));

     loop->_tail = back_control = ift2;

     set_loop(ift2, loop);

     set_loop(iff2, get_loop(iffalse));

     // Lazy update of 'get_ctrl' mechanism.

     lazy_replace_proj( iffalse, iff2 );

     lazy_replace_proj( iftrue,  ift2 );

     // Swap names

     iffalse = iff2;

     iftrue  = ift2;

   } else {

     _igvn.hash_delete(iffalse);

     _igvn.hash_delete(iftrue);

     iffalse->set_req_X( 0, le, &_igvn );

     iftrue ->set_req_X( 0, le, &_igvn );

   }

   set_idom(iftrue,  le, dd+1);

   set_idom(iffalse, le, dd+1);

   assert(iff->outcnt() == 0, "should be dead now");

   lazy_replace( iff, le ); // fix 'get_ctrl'

   // Now setup a new CountedLoopNode to replace the existing LoopNode

   CountedLoopNode *l = new (C) CountedLoopNode(init_control, back_control);

   l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve

   // The following assert is approximately true, and defines the intention

   // of can_be_counted_loop.  It fails, however, because phase->type

   // is not yet initialized for this loop and its parts.

   //assert(l->can_be_counted_loop(this), "sanity");

   _igvn.register_new_node_with_optimizer(l);

   set_loop(l, loop);

   loop->_head = l;

   // Fix all data nodes placed at the old loop head.

   // Uses the lazy-update mechanism of 'get_ctrl'.

   lazy_replace( x, l );

   set_idom(l, init_control, dom_depth(x));

   // Check for immediately preceding SafePoint and remove

   Node *sfpt2 = le->in(0);

   if (sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2)) {

     lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));

     if (loop->_safepts != NULL) {

       loop->_safepts->yank(sfpt2);

     }

   }

   // Free up intermediate goo

   _igvn.remove_dead_node(hook);

 #ifdef ASSERT

   assert(l->is_valid_counted_loop(), "counted loop shape is messed up");

   assert(l == loop->_head && l->phi() == phi && l->loopexit() == lex, "" );

 #endif

 #ifndef PRODUCT

   if (TraceLoopOpts) {

     tty->print("Counted      ");

     loop->dump_head();

   }

 #endif

   C->print_method(PHASE_AFTER_CLOOPS, 3);

   return true;

 }

 //----------------------exact_limit-------------------------------------------

 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {

   assert(loop->_head->is_CountedLoop(), "");

   CountedLoopNode *cl = loop->_head->as_CountedLoop();

   assert(cl->is_valid_counted_loop(), "");

   if (!LoopLimitCheck || ABS(cl->stride_con()) == 1 ||

       cl->limit()->Opcode() == Op_LoopLimit) {

     // Old code has exact limit (it could be incorrect in case of int overflow).

     // Loop limit is exact with stride == 1. And loop may already have exact limit.

     return cl->limit();

   }

   Node *limit = NULL;

 #ifdef ASSERT

   BoolTest::mask bt = cl->loopexit()->test_trip();

   assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");

 #endif

   if (cl->has_exact_trip_count()) {

     // Simple case: loop has constant boundaries.

     // Use jlongs to avoid integer overflow.

     int stride_con = cl->stride_con();

     jlong  init_con = cl->init_trip()->get_int();

     jlong limit_con = cl->limit()->get_int();

     julong trip_cnt = cl->trip_count();

     jlong final_con = init_con + trip_cnt*stride_con;

     int final_int = (int)final_con;

     // The final value should be in integer range since the loop

     // is counted and the limit was checked for overflow.

     assert(final_con == (jlong)final_int, "final value should be integer");

     limit = _igvn.intcon(final_int);

   } else {

     // Create new LoopLimit node to get exact limit (final iv value).

     limit = new (C) LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());

     register_new_node(limit, cl->in(LoopNode::EntryControl));

   }

   assert(limit != NULL, "sanity");

   return limit;

 }

 //------------------------------Ideal------------------------------------------

 // Return a node which is more "ideal" than the current node.

 // Attempt to convert into a counted-loop.

 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {

   if (!can_be_counted_loop(phase)) {

     phase->C->set_major_progress();

   }

   return RegionNode::Ideal(phase, can_reshape);

 }

 //=============================================================================

 //------------------------------Ideal------------------------------------------

 // Return a node which is more "ideal" than the current node.

 // Attempt to convert into a counted-loop.

 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {

   return RegionNode::Ideal(phase, can_reshape);

 }

 //------------------------------dump_spec--------------------------------------

 // Dump special per-node info

 #ifndef PRODUCT

 void CountedLoopNode::dump_spec(outputStream *st) const {

   LoopNode::dump_spec(st);

   if (stride_is_con()) {

     st->print("stride: %d ",stride_con());

   }

   if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);

   if (is_main_loop()) st->print("main of N%d", _idx);

   if (is_post_loop()) st->print("post of N%d", _main_idx);

 }

 #endif

 //=============================================================================

 int CountedLoopEndNode::stride_con() const {

   return stride()->bottom_type()->is_int()->get_con();

 }

 //=============================================================================

 //------------------------------Value-----------------------------------------

 const Type *LoopLimitNode::Value( PhaseTransform *phase ) const {

   const Type* init_t   = phase->type(in(Init));

   const Type* limit_t  = phase->type(in(Limit));

   const Type* stride_t = phase->type(in(Stride));

   // Either input is TOP ==> the result is TOP

   if (init_t   == Type::TOP) return Type::TOP;

   if (limit_t  == Type::TOP) return Type::TOP;

   if (stride_t == Type::TOP) return Type::TOP;

   int stride_con = stride_t->is_int()->get_con();

   if (stride_con == 1)

     return NULL;  // Identity

   if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {

     // Use jlongs to avoid integer overflow.

     jlong init_con   =  init_t->is_int()->get_con();

     jlong limit_con  = limit_t->is_int()->get_con();

     int  stride_m   = stride_con - (stride_con > 0 ? 1 : -1);

     jlong trip_count = (limit_con - init_con + stride_m)/stride_con;

     jlong final_con  = init_con + stride_con*trip_count;

     int final_int = (int)final_con;

     // The final value should be in integer range since the loop

     // is counted and the limit was checked for overflow.

     assert(final_con == (jlong)final_int, "final value should be integer");

     return TypeInt::make(final_int);

   }

   return bottom_type(); // TypeInt::INT

 }

 //------------------------------Ideal------------------------------------------

 // Return a node which is more "ideal" than the current node.

 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {

   if (phase->type(in(Init))   == Type::TOP ||

       phase->type(in(Limit))  == Type::TOP ||

       phase->type(in(Stride)) == Type::TOP)

     return NULL;  // Dead

   int stride_con = phase->type(in(Stride))->is_int()->get_con();

   if (stride_con == 1)

     return NULL;  // Identity

   if (in(Init)->is_Con() && in(Limit)->is_Con())

     return NULL;  // Value

   // Delay following optimizations until all loop optimizations

   // done to keep Ideal graph simple.

   if (!can_reshape || phase->C->major_progress())

     return NULL;

   const TypeInt* init_t  = phase->type(in(Init) )->is_int();

   const TypeInt* limit_t = phase->type(in(Limit))->is_int();

   int stride_p;

   jlong lim, ini;

   julong max;

   if (stride_con > 0) {

     stride_p = stride_con;

     lim = limit_t->_hi;

     ini = init_t->_lo;

     max = (julong)max_jint;

   } else {

     stride_p = -stride_con;

     lim = init_t->_hi;

     ini = limit_t->_lo;

     max = (julong)min_jint;

   }

   julong range = lim - ini + stride_p;

   if (range <= max) {

     // Convert to integer expression if it is not overflow.

     Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));

     Node *range = phase->transform(new (phase->C) SubINode(in(Limit), in(Init)));

     Node *bias  = phase->transform(new (phase->C) AddINode(range, stride_m));

     Node *trip  = phase->transform(new (phase->C) DivINode(0, bias, in(Stride)));

     Node *span  = phase->transform(new (phase->C) MulINode(trip, in(Stride)));

     return new (phase->C) AddINode(span, in(Init)); // exact limit

   }

   if (is_power_of_2(stride_p) ||                // divisor is 2^n

       !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?

     // Convert to long expression to avoid integer overflow

     // and let igvn optimizer convert this division.

     //

     Node*   init   = phase->transform( new (phase->C) ConvI2LNode(in(Init)));

     Node*  limit   = phase->transform( new (phase->C) ConvI2LNode(in(Limit)));

     Node* stride   = phase->longcon(stride_con);

     Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));

     Node *range = phase->transform(new (phase->C) SubLNode(limit, init));

     Node *bias  = phase->transform(new (phase->C) AddLNode(range, stride_m));

     Node *span;

     if (stride_con > 0 && is_power_of_2(stride_p)) {

       // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)

       // and avoid generating rounding for division. Zero trip guard should

       // guarantee that init < limit but sometimes the guard is missing and

       // we can get situation when init > limit. Note, for the empty loop

       // optimization zero trip guard is generated explicitly which leaves

       // only RCE predicate where exact limit is used and the predicate

       // will simply fail forcing recompilation.

       Node* neg_stride   = phase->longcon(-stride_con);

       span = phase->transform(new (phase->C) AndLNode(bias, neg_stride));

     } else {

       Node *trip  = phase->transform(new (phase->C) DivLNode(0, bias, stride));

       span = phase->transform(new (phase->C) MulLNode(trip, stride));

     }

     // Convert back to int

     Node *span_int = phase->transform(new (phase->C) ConvL2INode(span));

     return new (phase->C) AddINode(span_int, in(Init)); // exact limit

   }

   return NULL;    // No progress

 }

 //------------------------------Identity---------------------------------------

 // If stride == 1 return limit node.

 Node *LoopLimitNode::Identity( PhaseTransform *phase ) {

   int stride_con = phase->type(in(Stride))->is_int()->get_con();

   if (stride_con == 1 || stride_con == -1)

     return in(Limit);

   return this;

 }

 //=============================================================================

 //----------------------match_incr_with_optional_truncation--------------------

 // Match increment with optional truncation:

 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16

 // Return NULL for failure. Success returns the increment node.

 Node* CountedLoopNode::match_incr_with_optional_truncation(

                       Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {

   // Quick cutouts:

   if (expr == NULL || expr->req() != 3)  return NULL;

   Node *t1 = NULL;

   Node *t2 = NULL;

   const TypeInt* trunc_t = TypeInt::INT;

   Node* n1 = expr;

   int   n1op = n1->Opcode();

   // Try to strip (n1 & M) or (n1 << N >> N) from n1.

   if (n1op == Op_AndI &&

       n1->in(2)->is_Con() &&

       n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {

     // %%% This check should match any mask of 2**K-1.

     t1 = n1;

     n1 = t1->in(1);

     n1op = n1->Opcode();

     trunc_t = TypeInt::CHAR;

   } else if (n1op == Op_RShiftI &&

              n1->in(1) != NULL &&

              n1->in(1)->Opcode() == Op_LShiftI &&

              n1->in(2) == n1->in(1)->in(2) &&

              n1->in(2)->is_Con()) {

     jint shift = n1->in(2)->bottom_type()->is_int()->get_con();

     // %%% This check should match any shift in [1..31].

     if (shift == 16 || shift == 8) {

       t1 = n1;

       t2 = t1->in(1);

       n1 = t2->in(1);

       n1op = n1->Opcode();

       if (shift == 16) {

         trunc_t = TypeInt::SHORT;

       } else if (shift == 8) {

         trunc_t = TypeInt::BYTE;

       }

     }

   }

   // If (maybe after stripping) it is an AddI, we won:

   if (n1op == Op_AddI) {

     *trunc1 = t1;

     *trunc2 = t2;

     *trunc_type = trunc_t;

     return n1;

   }

   // failed

   return NULL;

 }

 //------------------------------filtered_type--------------------------------

 // Return a type based on condition control flow

 // A successful return will be a type that is restricted due

 // to a series of dominating if-tests, such as:

 //    if (i < 10) {

 //       if (i > 0) {

 //          here: "i" type is [1..10)

 //       }

 //    }

 // or a control flow merge

 //    if (i < 10) {

 //       do {

 //          phi( , ) -- at top of loop type is [min_int..10)

 //         i = ?

 //       } while ( i < 10)

 //

 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {

   assert(n && n->bottom_type()->is_int(), "must be int");

   const TypeInt* filtered_t = NULL;

   if (!n->is_Phi()) {

     assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");

     filtered_t = filtered_type_from_dominators(n, n_ctrl);

   } else {

     Node* phi    = n->as_Phi();

     Node* region = phi->in(0);

     assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");

     if (region && region != C->top()) {

       for (uint i = 1; i < phi->req(); i++) {

         Node* val   = phi->in(i);

         Node* use_c = region->in(i);

         const TypeInt* val_t = filtered_type_from_dominators(val, use_c);

         if (val_t != NULL) {

           if (filtered_t == NULL) {

             filtered_t = val_t;

           } else {

             filtered_t = filtered_t->meet(val_t)->is_int();

           }

         }

       }

     }

   }

   const TypeInt* n_t = _igvn.type(n)->is_int();

   if (filtered_t != NULL) {

     n_t = n_t->join(filtered_t)->is_int();

   }

   return n_t;

 }

 //------------------------------filtered_type_from_dominators--------------------------------

 // Return a possibly more restrictive type for val based on condition control flow of dominators

 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {

   if (val->is_Con()) {

      return val->bottom_type()->is_int();

   }

   uint if_limit = 10; // Max number of dominating if's visited

   const TypeInt* rtn_t = NULL;

   if (use_ctrl && use_ctrl != C->top()) {

     Node* val_ctrl = get_ctrl(val);

     uint val_dom_depth = dom_depth(val_ctrl);

     Node* pred = use_ctrl;

     uint if_cnt = 0;

     while (if_cnt < if_limit) {

       if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {

         if_cnt++;

         const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);

         if (if_t != NULL) {

           if (rtn_t == NULL) {

             rtn_t = if_t;

           } else {

             rtn_t = rtn_t->join(if_t)->is_int();

           }

         }

       }

       pred = idom(pred);

       if (pred == NULL || pred == C->top()) {

         break;

       }

       // Stop if going beyond definition block of val

       if (dom_depth(pred) < val_dom_depth) {

         break;

       }

     }

   }

   return rtn_t;

 }

 //------------------------------dump_spec--------------------------------------

 // Dump special per-node info

 #ifndef PRODUCT

 void CountedLoopEndNode::dump_spec(outputStream *st) const {

   if( in(TestValue)->is_Bool() ) {

     BoolTest bt( test_trip()); // Added this for g++.

     st->print("[");

     bt.dump_on(st);

     st->print("]");

   }

   st->print(" ");

   IfNode::dump_spec(st);

 }

 #endif

 //=============================================================================

 //------------------------------is_member--------------------------------------

 // Is 'l' a member of 'this'?

 int IdealLoopTree::is_member( const IdealLoopTree *l ) const {

   while( l->_nest > _nest ) l = l->_parent;

   return l == this;

 }

 //------------------------------set_nest---------------------------------------

 // Set loop tree nesting depth.  Accumulate _has_call bits.

 int IdealLoopTree::set_nest( uint depth ) {

   _nest = depth;

   int bits = _has_call;

   if( _child ) bits |= _child->set_nest(depth+1);

   if( bits ) _has_call = 1;

   if( _next  ) bits |= _next ->set_nest(depth  );

   return bits;

 }

 //------------------------------split_fall_in----------------------------------

 // Split out multiple fall-in edges from the loop header.  Move them to a

 // private RegionNode before the loop.  This becomes the loop landing pad.

 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {

   PhaseIterGVN &igvn = phase->_igvn;

   uint i;

   // Make a new RegionNode to be the landing pad.

   Node *landing_pad = new (phase->C) RegionNode( fall_in_cnt+1 );

   phase->set_loop(landing_pad,_parent);

   // Gather all the fall-in control paths into the landing pad

   uint icnt = fall_in_cnt;

   uint oreq = _head->req();

   for( i = oreq-1; i>0; i-- )

     if( !phase->is_member( this, _head->in(i) ) )

       landing_pad->set_req(icnt--,_head->in(i));

   // Peel off PhiNode edges as well

   for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {

     Node *oj = _head->fast_out(j);

     if( oj->is_Phi() ) {

       PhiNode* old_phi = oj->as_Phi();

       assert( old_phi->region() == _head, "" );

       igvn.hash_delete(old_phi);   // Yank from hash before hacking edges

       Node *p = PhiNode::make_blank(landing_pad, old_phi);

       uint icnt = fall_in_cnt;

       for( i = oreq-1; i>0; i-- ) {

         if( !phase->is_member( this, _head->in(i) ) ) {

           p->init_req(icnt--, old_phi->in(i));

           // Go ahead and clean out old edges from old phi

           old_phi->del_req(i);

         }

       }

       // Search for CSE's here, because ZKM.jar does a lot of

       // loop hackery and we need to be a little incremental

       // with the CSE to avoid O(N^2) node blow-up.

       Node *p2 = igvn.hash_find_insert(p); // Look for a CSE

       if( p2 ) {                // Found CSE

         p->destruct();          // Recover useless new node

         p = p2;                 // Use old node

       } else {

         igvn.register_new_node_with_optimizer(p, old_phi);

       }

       // Make old Phi refer to new Phi.

       old_phi->add_req(p);

       // Check for the special case of making the old phi useless and

       // disappear it.  In JavaGrande I have a case where this useless

       // Phi is the loop limit and prevents recognizing a CountedLoop

       // which in turn prevents removing an empty loop.

       Node *id_old_phi = old_phi->Identity( &igvn );

       if( id_old_phi != old_phi ) { // Found a simple identity?

         // Note that I cannot call 'replace_node' here, because

         // that will yank the edge from old_phi to the Region and

         // I'm mid-iteration over the Region's uses.

         for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {

           Node* use = old_phi->last_out(i);

           igvn.rehash_node_delayed(use);

           uint uses_found = 0;

           for (uint j = 0; j < use->len(); j++) {

             if (use->in(j) == old_phi) {

               if (j < use->req()) use->set_req (j, id_old_phi);

               else                use->set_prec(j, id_old_phi);

               uses_found++;

             }

           }

           i -= uses_found;    // we deleted 1 or more copies of this edge

         }

       }

       igvn._worklist.push(old_phi);

     }

   }

   // Finally clean out the fall-in edges from the RegionNode

   for( i = oreq-1; i>0; i-- ) {

     if( !phase->is_member( this, _head->in(i) ) ) {

       _head->del_req(i);

     }

   }

   // Transform landing pad

   igvn.register_new_node_with_optimizer(landing_pad, _head);

   // Insert landing pad into the header

   _head->add_req(landing_pad);

 }

 //------------------------------split_outer_loop-------------------------------

 // Split out the outermost loop from this shared header.

 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {

   PhaseIterGVN &igvn = phase->_igvn;

   // Find index of outermost loop; it should also be my tail.

   uint outer_idx = 1;

   while( _head->in(outer_idx) != _tail ) outer_idx++;

   // Make a LoopNode for the outermost loop.

   Node *ctl = _head->in(LoopNode::EntryControl);

   Node *outer = new (phase->C) LoopNode( ctl, _head->in(outer_idx) );

   outer = igvn.register_new_node_with_optimizer(outer, _head);

   phase->set_created_loop_node();

   // Outermost loop falls into '_head' loop

   _head->set_req(LoopNode::EntryControl, outer);

   _head->del_req(outer_idx);

   // Split all the Phis up between '_head' loop and 'outer' loop.

   for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {

     Node *out = _head->fast_out(j);

     if( out->is_Phi() ) {

       PhiNode *old_phi = out->as_Phi();

       assert( old_phi->region() == _head, "" );

       Node *phi = PhiNode::make_blank(outer, old_phi);

       phi->init_req(LoopNode::EntryControl,    old_phi->in(LoopNode::EntryControl));

       phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));

       phi = igvn.register_new_node_with_optimizer(phi, old_phi);

       // Make old Phi point to new Phi on the fall-in path

       igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);

       old_phi->del_req(outer_idx);

     }

   }

   // Use the new loop head instead of the old shared one

   _head = outer;

   phase->set_loop(_head, this);

 }

 //------------------------------fix_parent-------------------------------------

 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {

   loop->_parent = parent;

   if( loop->_child ) fix_parent( loop->_child, loop   );

   if( loop->_next  ) fix_parent( loop->_next , parent );

 }

 //------------------------------estimate_path_freq-----------------------------

 static float estimate_path_freq( Node *n ) {

   // Try to extract some path frequency info

   IfNode *iff;

   for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests

     uint nop = n->Opcode();

     if( nop == Op_SafePoint ) {   // Skip any safepoint

       n = n->in(0);

       continue;

     }

     if( nop == Op_CatchProj ) {   // Get count from a prior call

       // Assume call does not always throw exceptions: means the call-site

       // count is also the frequency of the fall-through path.

       assert( n->is_CatchProj(), "" );

       if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )

         return 0.0f;            // Assume call exception path is rare

       Node *call = n->in(0)->in(0)->in(0);

       assert( call->is_Call(), "expect a call here" );

       const JVMState *jvms = ((CallNode*)call)->jvms();

       ciMethodData* methodData = jvms->method()->method_data();

       if (!methodData->is_mature())  return 0.0f; // No call-site data

       ciProfileData* data = methodData->bci_to_data(jvms->bci());

       if ((data == NULL) || !data->is_CounterData()) {

         // no call profile available, try call's control input

         n = n->in(0);

         continue;

       }

       return data->as_CounterData()->count()/FreqCountInvocations;

     }

     // See if there's a gating IF test

     Node *n_c = n->in(0);

     if( !n_c->is_If() ) break;       // No estimate available

     iff = n_c->as_If();

     if( iff->_fcnt != COUNT_UNKNOWN )   // Have a valid count?

       // Compute how much count comes on this path

       return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;

     // Have no count info.  Skip dull uncommon-trap like branches.

     if( (nop == Op_IfTrue  && iff->_prob < PROB_LIKELY_MAG(5)) ||

         (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )

       break;

     // Skip through never-taken branch; look for a real loop exit.

     n = iff->in(0);

   }

   return 0.0f;                  // No estimate available

 }

 //------------------------------merge_many_backedges---------------------------

 // Merge all the backedges from the shared header into a private Region.

 // Feed that region as the one backedge to this loop.

 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {

   uint i;

   // Scan for the top 2 hottest backedges

   float hotcnt = 0.0f;

   float warmcnt = 0.0f;

   uint hot_idx = 0;

   // Loop starts at 2 because slot 1 is the fall-in path

   for( i = 2; i < _head->req(); i++ ) {

     float cnt = estimate_path_freq(_head->in(i));

     if( cnt > hotcnt ) {       // Grab hottest path

       warmcnt = hotcnt;

       hotcnt = cnt;

       hot_idx = i;

     } else if( cnt > warmcnt ) { // And 2nd hottest path

       warmcnt = cnt;

     }

   }

   // See if the hottest backedge is worthy of being an inner loop

   // by being much hotter than the next hottest backedge.

   if( hotcnt <= 0.0001 ||

       hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge

   // Peel out the backedges into a private merge point; peel

   // them all except optionally hot_idx.

   PhaseIterGVN &igvn = phase->_igvn;

   Node *hot_tail = NULL;

   // Make a Region for the merge point

   Node *r = new (phase->C) RegionNode(1);

   for( i = 2; i < _head->req(); i++ ) {

     if( i != hot_idx )

       r->add_req( _head->in(i) );

     else hot_tail = _head->in(i);

   }

   igvn.register_new_node_with_optimizer(r, _head);

   // Plug region into end of loop _head, followed by hot_tail

   while( _head->req() > 3 ) _head->del_req( _head->req()-1 );

   _head->set_req(2, r);

   if( hot_idx ) _head->add_req(hot_tail);

   // Split all the Phis up between '_head' loop and the Region 'r'

   for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {

     Node *out = _head->fast_out(j);

     if( out->is_Phi() ) {

       PhiNode* n = out->as_Phi();

       igvn.hash_delete(n);      // Delete from hash before hacking edges

       Node *hot_phi = NULL;

       Node *phi = new (phase->C) PhiNode(r, n->type(), n->adr_type());

       // Check all inputs for the ones to peel out

       uint j = 1;

       for( uint i = 2; i < n->req(); i++ ) {

         if( i != hot_idx )

           phi->set_req( j++, n->in(i) );

         else hot_phi = n->in(i);

       }

       // Register the phi but do not transform until whole place transforms

       igvn.register_new_node_with_optimizer(phi, n);

       // Add the merge phi to the old Phi

       while( n->req() > 3 ) n->del_req( n->req()-1 );

       n->set_req(2, phi);

       if( hot_idx ) n->add_req(hot_phi);

     }

   }

   // Insert a new IdealLoopTree inserted below me.  Turn it into a clone

   // of self loop tree.  Turn self into a loop headed by _head and with

   // tail being the new merge point.

   IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );

   phase->set_loop(_tail,ilt);   // Adjust tail

   _tail = r;                    // Self's tail is new merge point

   phase->set_loop(r,this);

   ilt->_child = _child;         // New guy has my children

   _child = ilt;                 // Self has new guy as only child

   ilt->_parent = this;          // new guy has self for parent

   ilt->_nest = _nest;           // Same nesting depth (for now)

   // Starting with 'ilt', look for child loop trees using the same shared

   // header.  Flatten these out; they will no longer be loops in the end.

   IdealLoopTree **pilt = &_child;

   while( ilt ) {

     if( ilt->_head == _head ) {

       uint i;

       for( i = 2; i < _head->req(); i++ )

         if( _head->in(i) == ilt->_tail )

           break;                // Still a loop

       if( i == _head->req() ) { // No longer a loop

         // Flatten ilt.  Hang ilt's "_next" list from the end of

         // ilt's '_child' list.  Move the ilt's _child up to replace ilt.

         IdealLoopTree **cp = &ilt->_child;

         while( *cp ) cp = &(*cp)->_next;   // Find end of child list

         *cp = ilt->_next;       // Hang next list at end of child list

         *pilt = ilt->_child;    // Move child up to replace ilt

         ilt->_head = NULL;      // Flag as a loop UNIONED into parent

         ilt = ilt->_child;      // Repeat using new ilt

         continue;               // do not advance over ilt->_child

       }

       assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );

       phase->set_loop(_head,ilt);

     }

     pilt = &ilt->_child;        // Advance to next

     ilt = *pilt;

   }

   if( _child ) fix_parent( _child, this );

 }

 //------------------------------beautify_loops---------------------------------

 // Split shared headers and insert loop landing pads.

 // Insert a LoopNode to replace the RegionNode.

 // Return TRUE if loop tree is structurally changed.

 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {

   bool result = false;

   // Cache parts in locals for easy

   PhaseIterGVN &igvn = phase->_igvn;

   igvn.hash_delete(_head);      // Yank from hash before hacking edges

   // Check for multiple fall-in paths.  Peel off a landing pad if need be.

   int fall_in_cnt = 0;

   for( uint i = 1; i < _head->req(); i++ )

     if( !phase->is_member( this, _head->in(i) ) )

       fall_in_cnt++;

   assert( fall_in_cnt, "at least 1 fall-in path" );

   if( fall_in_cnt > 1 )         // Need a loop landing pad to merge fall-ins

     split_fall_in( phase, fall_in_cnt );

   // Swap inputs to the _head and all Phis to move the fall-in edge to

   // the left.

   fall_in_cnt = 1;

   while( phase->is_member( this, _head->in(fall_in_cnt) ) )

     fall_in_cnt++;

   if( fall_in_cnt > 1 ) {

     // Since I am just swapping inputs I do not need to update def-use info

     Node *tmp = _head->in(1);

     _head->set_req( 1, _head->in(fall_in_cnt) );

     _head->set_req( fall_in_cnt, tmp );

     // Swap also all Phis

     for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {

       Node* phi = _head->fast_out(i);

       if( phi->is_Phi() ) {

         igvn.hash_delete(phi); // Yank from hash before hacking edges

         tmp = phi->in(1);

         phi->set_req( 1, phi->in(fall_in_cnt) );

         phi->set_req( fall_in_cnt, tmp );

       }

     }

   }

   assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );

   assert(  phase->is_member( this, _head->in(2) ), "right edge is loop" );

   // If I am a shared header (multiple backedges), peel off the many

   // backedges into a private merge point and use the merge point as

   // the one true backedge.

   if( _head->req() > 3 ) {

     // Merge the many backedges into a single backedge but leave

     // the hottest backedge as separate edge for the following peel.

     merge_many_backedges( phase );

     result = true;

   }

   // If I have one hot backedge, peel off myself loop.

   // I better be the outermost loop.

   if (_head->req() > 3 && !_irreducible) {

     split_outer_loop( phase );

     result = true;

   } else if (!_head->is_Loop() && !_irreducible) {

     // Make a new LoopNode to replace the old loop head

     Node *l = new (phase->C) LoopNode( _head->in(1), _head->in(2) );

     l = igvn.register_new_node_with_optimizer(l, _head);

     phase->set_created_loop_node();

     // Go ahead and replace _head

     phase->_igvn.replace_node( _head, l );

     _head = l;

     phase->set_loop(_head, this);

   }

   // Now recursively beautify nested loops

   if( _child ) result |= _child->beautify_loops( phase );

   if( _next  ) result |= _next ->beautify_loops( phase );

   return result;

 }

 //------------------------------allpaths_check_safepts----------------------------

 // Allpaths backwards scan from loop tail, terminating each path at first safepoint

 // encountered.  Helper for check_safepts.

 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {

   assert(stack.size() == 0, "empty stack");

   stack.push(_tail);

   visited.Clear();

   visited.set(_tail->_idx);

   while (stack.size() > 0) {

     Node* n = stack.pop();

     if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {

       // Terminate this path

     } else if (n->Opcode() == Op_SafePoint) {

       if (_phase->get_loop(n) != this) {

         if (_required_safept == NULL) _required_safept = new Node_List();

         _required_safept->push(n);  // save the one closest to the tail

       }

       // Terminate this path

     } else {

       uint start = n->is_Region() ? 1 : 0;

       uint end   = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;

       for (uint i = start; i < end; i++) {

         Node* in = n->in(i);

         assert(in->is_CFG(), "must be");

         if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {

           stack.push(in);

         }

       }

     }

   }

 }

 //------------------------------check_safepts----------------------------

 // Given dominators, try to find loops with calls that must always be

 // executed (call dominates loop tail).  These loops do not need non-call

 // safepoints (ncsfpt).

 //

 // A complication is that a safepoint in a inner loop may be needed

 // by an outer loop. In the following, the inner loop sees it has a

 // call (block 3) on every path from the head (block 2) to the

 // backedge (arc 3->2).  So it deletes the ncsfpt (non-call safepoint)

 // in block 2, _but_ this leaves the outer loop without a safepoint.

 //

 //          entry  0

 //                 |

 //                 v

 // outer 1,2    +->1

 //              |  |

 //              |  v

 //              |  2<---+  ncsfpt in 2

 //              |_/|\   |

 //                 | v  |

 // inner 2,3      /  3  |  call in 3

 //               /   |  |

 //              v    +--+

 //        exit  4

 //

 //

 // This method creates a list (_required_safept) of ncsfpt nodes that must

 // be protected is created for each loop. When a ncsfpt maybe deleted, it

 // is first looked for in the lists for the outer loops of the current loop.

 //

 // The insights into the problem:

 //  A) counted loops are okay

 //  B) innermost loops are okay (only an inner loop can delete

 //     a ncsfpt needed by an outer loop)

 //  C) a loop is immune from an inner loop deleting a safepoint

 //     if the loop has a call on the idom-path

 //  D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the

 //     idom-path that is not in a nested loop

 //  E) otherwise, an ncsfpt on the idom-path that is nested in an inner

 //     loop needs to be prevented from deletion by an inner loop

 //

 // There are two analyses:

 //  1) The first, and cheaper one, scans the loop body from

 //     tail to head following the idom (immediate dominator)

 //     chain, looking for the cases (C,D,E) above.

 //     Since inner loops are scanned before outer loops, there is summary

 //     information about inner loops.  Inner loops can be skipped over

 //     when the tail of an inner loop is encountered.

 //

 //  2) The second, invoked if the first fails to find a call or ncsfpt on

 //     the idom path (which is rare), scans all predecessor control paths

 //     from the tail to the head, terminating a path when a call or sfpt

 //     is encountered, to find the ncsfpt's that are closest to the tail.

 //

 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {

   // Bottom up traversal

   IdealLoopTree* ch = _child;

   if (_child) _child->check_safepts(visited, stack);

   if (_next)  _next ->check_safepts(visited, stack);

   if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {

     bool  has_call         = false; // call on dom-path

     bool  has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth

     Node* nonlocal_ncsfpt  = NULL;  // ncsfpt on dom-path at a deeper depth

     // Scan the dom-path nodes from tail to head

     for (Node* n = tail(); n != _head; n = _phase->idom(n)) {

       if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {

         has_call = true;

         _has_sfpt = 1;          // Then no need for a safept!

         break;

       } else if (n->Opcode() == Op_SafePoint) {

         if (_phase->get_loop(n) == this) {

           has_local_ncsfpt = true;

           break;

         }

         if (nonlocal_ncsfpt == NULL) {

           nonlocal_ncsfpt = n; // save the one closest to the tail

         }

       } else {

         IdealLoopTree* nlpt = _phase->get_loop(n);

         if (this != nlpt) {

           // If at an inner loop tail, see if the inner loop has already

           // recorded seeing a call on the dom-path (and stop.)  If not,

           // jump to the head of the inner loop.

           assert(is_member(nlpt), "nested loop");

           Node* tail = nlpt->_tail;

           if (tail->in(0)->is_If()) tail = tail->in(0);

           if (n == tail) {

             // If inner loop has call on dom-path, so does outer loop

             if (nlpt->_has_sfpt) {

               has_call = true;

               _has_sfpt = 1;

               break;

             }

             // Skip to head of inner loop

             assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");

             n = nlpt->_head;

           }

         }

       }

     }

     // Record safept's that this loop needs preserved when an

     // inner loop attempts to delete it's safepoints.

     if (_child != NULL && !has_call && !has_local_ncsfpt) {

       if (nonlocal_ncsfpt != NULL) {

         if (_required_safept == NULL) _required_safept = new Node_List();

         _required_safept->push(nonlocal_ncsfpt);

       } else {

         // Failed to find a suitable safept on the dom-path.  Now use

         // an all paths walk from tail to head, looking for safepoints to preserve.

         allpaths_check_safepts(visited, stack);

       }

     }

   }

 }

 //---------------------------is_deleteable_safept----------------------------

 // Is safept not required by an outer loop?

 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {

   assert(sfpt->Opcode() == Op_SafePoint, "");

   IdealLoopTree* lp = get_loop(sfpt)->_parent;

   while (lp != NULL) {

     Node_List* sfpts = lp->_required_safept;

     if (sfpts != NULL) {

       for (uint i = 0; i < sfpts->size(); i++) {

         if (sfpt == sfpts->at(i))

           return false;

       }

     }

     lp = lp->_parent;

   }

   return true;

 }

 //---------------------------replace_parallel_iv-------------------------------

 // Replace parallel induction variable (parallel to trip counter)

 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {

   assert(loop->_head->is_CountedLoop(), "");

   CountedLoopNode *cl = loop->_head->as_CountedLoop();

   if (!cl->is_valid_counted_loop())

     return;         // skip malformed counted loop

   Node *incr = cl->incr();

   if (incr == NULL)

     return;         // Dead loop?

   Node *init = cl->init_trip();

   Node *phi  = cl->phi();

   int stride_con = cl->stride_con();

   // Visit all children, looking for Phis

   for (DUIterator i = cl->outs(); cl->has_out(i); i++) {

     Node *out = cl->out(i);

     // Look for other phis (secondary IVs). Skip dead ones

     if (!out->is_Phi() || out == phi || !has_node(out))

       continue;

     PhiNode* phi2 = out->as_Phi();

     Node *incr2 = phi2->in( LoopNode::LoopBackControl );

     // Look for induction variables of the form:  X += constant

     if (phi2->region() != loop->_head ||

         incr2->req() != 3 ||

         incr2->in(1) != phi2 ||

         incr2 == incr ||

         incr2->Opcode() != Op_AddI ||

         !incr2->in(2)->is_Con())

       continue;

     // Check for parallel induction variable (parallel to trip counter)

     // via an affine function.  In particular, count-down loops with

     // count-up array indices are common. We only RCE references off

     // the trip-counter, so we need to convert all these to trip-counter

     // expressions.

     Node *init2 = phi2->in( LoopNode::EntryControl );

     int stride_con2 = incr2->in(2)->get_int();

     // The general case here gets a little tricky.  We want to find the

     // GCD of all possible parallel IV's and make a new IV using this

     // GCD for the loop.  Then all possible IVs are simple multiples of

     // the GCD.  In practice, this will cover very few extra loops.

     // Instead we require 'stride_con2' to be a multiple of 'stride_con',

     // where +/-1 is the common case, but other integer multiples are

     // also easy to handle.

     int ratio_con = stride_con2/stride_con;

     if ((ratio_con * stride_con) == stride_con2) { // Check for exact

 #ifndef PRODUCT

       if (TraceLoopOpts) {

         tty->print("Parallel IV: %d ", phi2->_idx);

         loop->dump_head();

       }

 #endif

       // Convert to using the trip counter.  The parallel induction

       // variable differs from the trip counter by a loop-invariant

       // amount, the difference between their respective initial values.

       // It is scaled by the 'ratio_con'.

       Node* ratio = _igvn.intcon(ratio_con);

       set_ctrl(ratio, C->root());

       Node* ratio_init = new (C) MulINode(init, ratio);

       _igvn.register_new_node_with_optimizer(ratio_init, init);

       set_early_ctrl(ratio_init);

       Node* diff = new (C) SubINode(init2, ratio_init);

       _igvn.register_new_node_with_optimizer(diff, init2);

       set_early_ctrl(diff);

       Node* ratio_idx = new (C) MulINode(phi, ratio);

       _igvn.register_new_node_with_optimizer(ratio_idx, phi);

       set_ctrl(ratio_idx, cl);

       Node* add = new (C) AddINode(ratio_idx, diff);

       _igvn.register_new_node_with_optimizer(add);

       set_ctrl(add, cl);

       _igvn.replace_node( phi2, add );

       // Sometimes an induction variable is unused

       if (add->outcnt() == 0) {

         _igvn.remove_dead_node(add);

       }

       --i; // deleted this phi; rescan starting with next position

       continue;

     }

   }

 }

 //------------------------------counted_loop-----------------------------------

 // Convert to counted loops where possible

 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {

   // For grins, set the inner-loop flag here

   if (!_child) {

     if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();

   }

   if (_head->is_CountedLoop() ||

       phase->is_counted_loop(_head, this)) {

     _has_sfpt = 1;              // Indicate we do not need a safepoint here

     // Look for safepoints to remove.

     Node_List* sfpts = _safepts;

     if (sfpts != NULL) {

       for (uint i = 0; i < sfpts->size(); i++) {

         Node* n = sfpts->at(i);

         assert(phase->get_loop(n) == this, "");

         if (phase->is_deleteable_safept(n)) {

           phase->lazy_replace(n, n->in(TypeFunc::Control));

         }

       }

     }

     // Look for induction variables

     phase->replace_parallel_iv(this);

   } else if (_parent != NULL && !_irreducible) {

     // Not a counted loop.

     // Look for a safepoint on the idom-path.

     Node* sfpt = tail();

     for (; sfpt != _head; sfpt = phase->idom(sfpt)) {

       if (sfpt->Opcode() == Op_SafePoint && phase->get_loop(sfpt) == this)

         break; // Found one

     }

     // Delete other safepoints in this loop.

     Node_List* sfpts = _safepts;

     if (sfpts != NULL && sfpt != _head && sfpt->Opcode() == Op_SafePoint) {

       for (uint i = 0; i < sfpts->size(); i++) {

         Node* n = sfpts->at(i);

         assert(phase->get_loop(n) == this, "");

         if (n != sfpt && phase->is_deleteable_safept(n)) {

           phase->lazy_replace(n, n->in(TypeFunc::Control));

         }

       }

     }

   }

   // Recursively

   if (_child) _child->counted_loop( phase );

   if (_next)  _next ->counted_loop( phase );

 }

 #ifndef PRODUCT

 //------------------------------dump_head--------------------------------------

 // Dump 1 liner for loop header info

 void IdealLoopTree::dump_head( ) const {

   for (uint i=0; i<_nest; i++)

     tty->print("  ");

   tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);

   if (_irreducible) tty->print(" IRREDUCIBLE");

   Node* entry = _head->in(LoopNode::EntryControl);

   if (LoopLimitCheck) {

     Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);

     if (predicate != NULL ) {

       tty->print(" limit_check");

       entry = entry->in(0)->in(0);

     }

   }

   if (UseLoopPredicate) {

     entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);

     if (entry != NULL) {

       tty->print(" predicated");

     }

   }

   if (_head->is_CountedLoop()) {

     CountedLoopNode *cl = _head->as_CountedLoop();

     tty->print(" counted");

     Node* init_n = cl->init_trip();

     if (init_n  != NULL &&  init_n->is_Con())

       tty->print(" [%d,", cl->init_trip()->get_int());

     else

       tty->print(" [int,");

     Node* limit_n = cl->limit();

     if (limit_n  != NULL &&  limit_n->is_Con())

       tty->print("%d),", cl->limit()->get_int());

     else

       tty->print("int),");

     int stride_con  = cl->stride_con();

     if (stride_con > 0) tty->print("+");

     tty->print("%d", stride_con);

     tty->print(" (%d iters) ", (int)cl->profile_trip_cnt());

     if (cl->is_pre_loop ()) tty->print(" pre" );

     if (cl->is_main_loop()) tty->print(" main");

     if (cl->is_post_loop()) tty->print(" post");

   }

   tty->cr();

 }

 //------------------------------dump-------------------------------------------

 // Dump loops by loop tree

 void IdealLoopTree::dump( ) const {

   dump_head();

   if (_child) _child->dump();

   if (_next)  _next ->dump();

 }

 #endif

 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {

   if (loop == root) {

     if (loop->_child != NULL) {

       log->begin_head("loop_tree");

       log->end_head();

       if( loop->_child ) log_loop_tree(root, loop->_child, log);

       log->tail("loop_tree");

       assert(loop->_next == NULL, "what?");

     }

   } else {

     Node* head = loop->_head;

     log->begin_head("loop");

     log->print(" idx='%d' ", head->_idx);

     if (loop->_irreducible) log->print("irreducible='1' ");

     if (head->is_Loop()) {

       if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");

       if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");

     }

     if (head->is_CountedLoop()) {

       CountedLoopNode* cl = head->as_CountedLoop();

       if (cl->is_pre_loop())  log->print("pre_loop='%d' ",  cl->main_idx());

       if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);

       if (cl->is_post_loop()) log->print("post_loop='%d' ",  cl->main_idx());

     }

     log->end_head();

     if( loop->_child ) log_loop_tree(root, loop->_child, log);

     log->tail("loop");

     if( loop->_next  ) log_loop_tree(root, loop->_next, log);

   }

 }

 //---------------------collect_potentially_useful_predicates-----------------------

 // Helper function to collect potentially useful predicates to prevent them from

 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates

 void PhaseIdealLoop::collect_potentially_useful_predicates(

                          IdealLoopTree * loop, Unique_Node_List &useful_predicates) {

   if (loop->_child) { // child

     collect_potentially_useful_predicates(loop->_child, useful_predicates);

   }

   // self (only loops that we can apply loop predication may use their predicates)

   if (loop->_head->is_Loop() &&

       !loop->_irreducible    &&

       !loop->tail()->is_top()) {

     LoopNode* lpn = loop->_head->as_Loop();

     Node* entry = lpn->in(LoopNode::EntryControl);

     Node* predicate_proj = find_predicate(entry); // loop_limit_check first

     if (predicate_proj != NULL ) { // right pattern that can be used by loop predication

       assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");

       useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one

       entry = entry->in(0)->in(0);

     }

     predicate_proj = find_predicate(entry); // Predicate

     if (predicate_proj != NULL ) {

       useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one

     }

   }

   if (loop->_next) { // sibling

     collect_potentially_useful_predicates(loop->_next, useful_predicates);

   }

 }

 //------------------------eliminate_useless_predicates-----------------------------

 // Eliminate all inserted predicates if they could not be used by loop predication.

 // Note: it will also eliminates loop limits check predicate since it also uses

 // Opaque1 node (see Parse::add_predicate()).

 void PhaseIdealLoop::eliminate_useless_predicates() {

   if (C->predicate_count() == 0)

     return; // no predicate left

   Unique_Node_List useful_predicates; // to store useful predicates

   if (C->has_loops()) {

     collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);

   }

   for (int i = C->predicate_count(); i > 0; i--) {

      Node * n = C->predicate_opaque1_node(i-1);

      assert(n->Opcode() == Op_Opaque1, "must be");

      if (!useful_predicates.member(n)) { // not in the useful list

        _igvn.replace_node(n, n->in(1));

      }

   }

 }

 //------------------------process_expensive_nodes-----------------------------

 // Expensive nodes have their control input set to prevent the GVN

 // from commoning them and as a result forcing the resulting node to

 // be in a more frequent path. Use CFG information here, to change the

 // control inputs so that some expensive nodes can be commoned while

 // not executed more frequently.

 bool PhaseIdealLoop::process_expensive_nodes() {

   assert(OptimizeExpensiveOps, "optimization off?");

   // Sort nodes to bring similar nodes together

   C->sort_expensive_nodes();

   bool progress = false;

   for (int i = 0; i < C->expensive_count(); ) {

     Node* n = C->expensive_node(i);

     int start = i;

     // Find nodes similar to n

     i++;

     for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);

     int end = i;

     // And compare them two by two

     for (int j = start; j < end; j++) {

       Node* n1 = C->expensive_node(j);

       if (is_node_unreachable(n1)) {

         continue;

       }

       for (int k = j+1; k < end; k++) {

         Node* n2 = C->expensive_node(k);

         if (is_node_unreachable(n2)) {

           continue;

         }

         assert(n1 != n2, "should be pair of nodes");

         Node* c1 = n1->in(0);

         Node* c2 = n2->in(0);

         Node* parent_c1 = c1;

         Node* parent_c2 = c2;

         // The call to get_early_ctrl_for_expensive() moves the

         // expensive nodes up but stops at loops that are in a if

         // branch. See whether we can exit the loop and move above the

         // If.

         if (c1->is_Loop()) {

           parent_c1 = c1->in(1);

         }

         if (c2->is_Loop()) {

           parent_c2 = c2->in(1);

         }

         if (parent_c1 == parent_c2) {

           _igvn._worklist.push(n1);

           _igvn._worklist.push(n2);

           continue;

         }

         // Look for identical expensive node up the dominator chain.

         if (is_dominator(c1, c2)) {

           c2 = c1;

         } else if (is_dominator(c2, c1)) {

           c1 = c2;

         } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&

                    parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {

           // Both branches have the same expensive node so move it up

           // before the if.

           c1 = c2 = idom(parent_c1->in(0));

         }

         // Do the actual moves

         if (n1->in(0) != c1) {

           _igvn.hash_delete(n1);

           n1->set_req(0, c1);

           _igvn.hash_insert(n1);

           _igvn._worklist.push(n1);

           progress = true;

         }

         if (n2->in(0) != c2) {

           _igvn.hash_delete(n2);

           n2->set_req(0, c2);

           _igvn.hash_insert(n2);

           _igvn._worklist.push(n2);

           progress = true;

         }

       }

     }

   }

   return progress;

 }

 //=============================================================================

 //----------------------------build_and_optimize-------------------------------

 // Create a PhaseLoop.  Build the ideal Loop tree.  Map each Ideal Node to

 // its corresponding LoopNode.  If 'optimize' is true, do some loop cleanups.

 void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool skip_loop_opts) {

   ResourceMark rm;

   int old_progress = C->major_progress();

   uint orig_worklist_size = _igvn._worklist.size();

   // Reset major-progress flag for the driver's heuristics

   C->clear_major_progress();

 #ifndef PRODUCT

   // Capture for later assert

   uint unique = C->unique();

   _loop_invokes++;

   _loop_work += unique;

 #endif

   // True if the method has at least 1 irreducible loop

   _has_irreducible_loops = false;

   _created_loop_node = false;

   Arena *a = Thread::current()->resource_area();

   VectorSet visited(a);

   // Pre-grow the mapping from Nodes to IdealLoopTrees.

   _nodes.map(C->unique(), NULL);

   memset(_nodes.adr(), 0, wordSize * C->unique());

   // Pre-build the top-level outermost loop tree entry

   _ltree_root = new IdealLoopTree( this, C->root(), C->root() );

   // Do not need a safepoint at the top level

   _ltree_root->_has_sfpt = 1;

   // Initialize Dominators.

   // Checked in clone_loop_predicate() during beautify_loops().

   _idom_size = 0;

   _idom      = NULL;

   _dom_depth = NULL;

   _dom_stk   = NULL;

   // Empty pre-order array

   allocate_preorders();

   // Build a loop tree on the fly.  Build a mapping from CFG nodes to

   // IdealLoopTree entries.  Data nodes are NOT walked.

   build_loop_tree();

   // Check for bailout, and return

   if (C->failing()) {

     return;

   }

   // No loops after all

   if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);

   // There should always be an outer loop containing the Root and Return nodes.

   // If not, we have a degenerate empty program.  Bail out in this case.

   if (!has_node(C->root())) {

     if (!_verify_only) {

       C->clear_major_progress();

       C->record_method_not_compilable("empty program detected during loop optimization");

     }

     return;

   }

   // Nothing to do, so get out

   bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only;

   bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);

   if (stop_early && !do_expensive_nodes) {

     _igvn.optimize();           // Cleanup NeverBranches

     return;

   }

   // Set loop nesting depth

   _ltree_root->set_nest( 0 );

   // Split shared headers and insert loop landing pads.

   // Do not bother doing this on the Root loop of course.

   if( !_verify_me && !_verify_only && _ltree_root->_child ) {

     C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);

     if( _ltree_root->_child->beautify_loops( this ) ) {

       // Re-build loop tree!

       _ltree_root->_child = NULL;

       _nodes.clear();

       reallocate_preorders();

       build_loop_tree();

       // Check for bailout, and return

       if (C->failing()) {

         return;

       }

       // Reset loop nesting depth

       _ltree_root->set_nest( 0 );

       C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);

     }

   }

   // Build Dominators for elision of NULL checks & loop finding.

   // Since nodes do not have a slot for immediate dominator, make

   // a persistent side array for that info indexed on node->_idx.

   _idom_size = C->unique();

   _idom      = NEW_RESOURCE_ARRAY( Node*, _idom_size );

   _dom_depth = NEW_RESOURCE_ARRAY( uint,  _idom_size );

   _dom_stk   = NULL; // Allocated on demand in recompute_dom_depth

   memset( _dom_depth, 0, _idom_size * sizeof(uint) );

   Dominators();

   if (!_verify_only) {

     // As a side effect, Dominators removed any unreachable CFG paths

     // into RegionNodes.  It doesn't do this test against Root, so

     // we do it here.

     for( uint i = 1; i < C->root()->req(); i++ ) {

       if( !_nodes[C->root()->in(i)->_idx] ) {    // Dead path into Root?

         _igvn.delete_input_of(C->root(), i);

         i--;                      // Rerun same iteration on compressed edges

       }

     }

     // Given dominators, try to find inner loops with calls that must

     // always be executed (call dominates loop tail).  These loops do

     // not need a separate safepoint.

     Node_List cisstack(a);

     _ltree_root->check_safepts(visited, cisstack);

   }

   // Walk the DATA nodes and place into loops.  Find earliest control

   // node.  For CFG nodes, the _nodes array starts out and remains

   // holding the associated IdealLoopTree pointer.  For DATA nodes, the

   // _nodes array holds the earliest legal controlling CFG node.

   // Allocate stack with enough space to avoid frequent realloc

   int stack_size = (C->unique() >> 1) + 16; // (unique>>1)+16 from Java2D stats

   Node_Stack nstack( a, stack_size );

   visited.Clear();

   Node_List worklist(a);

   // Don't need C->root() on worklist since

   // it will be processed among C->top() inputs

   worklist.push( C->top() );

   visited.set( C->top()->_idx ); // Set C->top() as visited now

   build_loop_early( visited, worklist, nstack );

   // Given early legal placement, try finding counted loops.  This placement

   // is good enough to discover most loop invariants.

   if( !_verify_me && !_verify_only )

     _ltree_root->counted_loop( this );

   // Find latest loop placement.  Find ideal loop placement.

   visited.Clear();

   init_dom_lca_tags();

   // Need C->root() on worklist when processing outs

   worklist.push( C->root() );

   NOT_PRODUCT( C->verify_graph_edges(); )

   worklist.push( C->top() );

   build_loop_late( visited, worklist, nstack );

   if (_verify_only) {

     // restore major progress flag

     for (int i = 0; i < old_progress; i++)

       C->set_major_progress();

     assert(C->unique() == unique, "verification mode made Nodes? ? ?");

     assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");

     return;

   }

   // clear out the dead code after build_loop_late

   while (_deadlist.size()) {

     _igvn.remove_globally_dead_node(_deadlist.pop());

   }

   if (stop_early) {

     assert(do_expensive_nodes, "why are we here?");

     if (process_expensive_nodes()) {

       // If we made some progress when processing expensive nodes then

       // the IGVN may modify the graph in a way that will allow us to

       // make some more progress: we need to try processing expensive

       // nodes again.

       C->set_major_progress();

     }

     _igvn.optimize();

     return;

   }

   // Some parser-inserted loop predicates could never be used by loop

   // predication or they were moved away from loop during some optimizations.

   // For example, peeling. Eliminate them before next loop optimizations.

   if (UseLoopPredicate || LoopLimitCheck) {

     eliminate_useless_predicates();

   }

 #ifndef PRODUCT

   C->verify_graph_edges();

   if (_verify_me) {             // Nested verify pass?

     // Check to see if the verify mode is broken

     assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");

     return;

   }

   if(VerifyLoopOptimizations) verify();

   if(TraceLoopOpts && C->has_loops()) {

     _ltree_root->dump();

   }

 #endif

   if (skip_loop_opts) {

     // Cleanup any modified bits

     _igvn.optimize();

     if (C->log() != NULL) {

       log_loop_tree(_ltree_root, _ltree_root, C->log());

     }

     return;

   }

   if (ReassociateInvariants) {

     // Reassociate invariants and prep for split_thru_phi

     for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {

       IdealLoopTree* lpt = iter.current();

       if (!lpt->is_counted() || !lpt->is_inner()) continue;

       lpt->reassociate_invariants(this);

       // Because RCE opportunities can be masked by split_thru_phi,

       // look for RCE candidates and inhibit split_thru_phi

       // on just their loop-phi's for this pass of loop opts

       if (SplitIfBlocks && do_split_ifs) {

         if (lpt->policy_range_check(this)) {

           lpt->_rce_candidate = 1; // = true

         }

       }

     }

   }

   // Check for aggressive application of split-if and other transforms

   // that require basic-block info (like cloning through Phi's)

   if( SplitIfBlocks && do_split_ifs ) {

     visited.Clear();

     split_if_with_blocks( visited, nstack );

     NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );

   }

   if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {

     C->set_major_progress();

   }

   // Perform loop predication before iteration splitting

   if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {

     _ltree_root->_child->loop_predication(this);

   }

   if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {

     if (do_intrinsify_fill()) {

       C->set_major_progress();

     }

   }

   // Perform iteration-splitting on inner loops.  Split iterations to avoid

   // range checks or one-shot null checks.

   // If split-if's didn't hack the graph too bad (no CFG changes)

   // then do loop opts.

   if (C->has_loops() && !C->major_progress()) {

     memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );

     _ltree_root->_child->iteration_split( this, worklist );

     // No verify after peeling!  GCM has hoisted code out of the loop.

     // After peeling, the hoisted code could sink inside the peeled area.

     // The peeling code does not try to recompute the best location for

     // all the code before the peeled area, so the verify pass will always

     // complain about it.

   }

   // Do verify graph edges in any case

   NOT_PRODUCT( C->verify_graph_edges(); );

   if (!do_split_ifs) {

     // We saw major progress in Split-If to get here.  We forced a

     // pass with unrolling and not split-if, however more split-if's

     // might make progress.  If the unrolling didn't make progress

     // then the major-progress flag got cleared and we won't try

     // another round of Split-If.  In particular the ever-common

     // instance-of/check-cast pattern requires at least 2 rounds of

     // Split-If to clear out.

     C->set_major_progress();

   }

   // Repeat loop optimizations if new loops were seen

   if (created_loop_node()) {

     C->set_major_progress();

   }

   // Keep loop predicates and perform optimizations with them

   // until no more loop optimizations could be done.

   // After that switch predicates off and do more loop optimizations.

   if (!C->major_progress() && (C->predicate_count() > 0)) {

      C->cleanup_loop_predicates(_igvn);

 #ifndef PRODUCT

      if (TraceLoopOpts) {

        tty->print_cr("PredicatesOff");

      }

 #endif

      C->set_major_progress();

   }

   // Convert scalar to superword operations at the end of all loop opts.

   if (UseSuperWord && C->has_loops() && !C->major_progress()) {

     // SuperWord transform

     SuperWord sw(this);

     for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {

       IdealLoopTree* lpt = iter.current();

       if (lpt->is_counted()) {

         sw.transform_loop(lpt);

       }

     }

   }

   // Cleanup any modified bits

   _igvn.optimize();

   // disable assert until issue with split_flow_path is resolved (6742111)

   // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),

   //        "shouldn't introduce irreducible loops");

   if (C->log() != NULL) {

     log_loop_tree(_ltree_root, _ltree_root, C->log());

   }

 }

 #ifndef PRODUCT

 //------------------------------print_statistics-------------------------------

 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes

 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique

 void PhaseIdealLoop::print_statistics() {

   tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);

 }

 //------------------------------verify-----------------------------------------

 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.

 static int fail;                // debug only, so its multi-thread dont care

 void PhaseIdealLoop::verify() const {

   int old_progress = C->major_progress();

   ResourceMark rm;

   PhaseIdealLoop loop_verify( _igvn, this );

   VectorSet visited(Thread::current()->resource_area());

   fail = 0;

   verify_compare( C->root(), &loop_verify, visited );

   assert( fail == 0, "verify loops failed" );

   // Verify loop structure is the same

   _ltree_root->verify_tree(loop_verify._ltree_root, NULL);

   // Reset major-progress.  It was cleared by creating a verify version of

   // PhaseIdealLoop.

   for( int i=0; i<old_progress; i++ )

     C->set_major_progress();

 }

 //------------------------------verify_compare---------------------------------

 // Make sure me and the given PhaseIdealLoop agree on key data structures

 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {

   if( !n ) return;

   if( visited.test_set( n->_idx ) ) return;

   if( !_nodes[n->_idx] ) {      // Unreachable

     assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );

     return;

   }

   uint i;

   for( i = 0; i < n->req(); i++ )

     verify_compare( n->in(i), loop_verify, visited );

   // Check the '_nodes' block/loop structure

   i = n->_idx;

   if( has_ctrl(n) ) {           // We have control; verify has loop or ctrl

     if( _nodes[i] != loop_verify->_nodes[i] &&

         get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {

       tty->print("Mismatched control setting for: ");

       n->dump();

       if( fail++ > 10 ) return;

       Node *c = get_ctrl_no_update(n);

       tty->print("We have it as: ");

       if( c->in(0) ) c->dump();

         else tty->print_cr("N%d",c->_idx);

       tty->print("Verify thinks: ");

       if( loop_verify->has_ctrl(n) )

         loop_verify->get_ctrl_no_update(n)->dump();

       else

         loop_verify->get_loop_idx(n)->dump();

       tty->cr();

     }

   } else {                    // We have a loop

     IdealLoopTree *us = get_loop_idx(n);

     if( loop_verify->has_ctrl(n) ) {

       tty->print("Mismatched loop setting for: ");

       n->dump();

       if( fail++ > 10 ) return;

       tty->print("We have it as: ");

       us->dump();

       tty->print("Verify thinks: ");

       loop_verify->get_ctrl_no_update(n)->dump();

       tty->cr();

     } else if (!C->major_progress()) {

       // Loop selection can be messed up if we did a major progress

       // operation, like split-if.  Do not verify in that case.

       IdealLoopTree *them = loop_verify->get_loop_idx(n);

       if( us->_head != them->_head ||  us->_tail != them->_tail ) {

         tty->print("Unequals loops for: ");

         n->dump();

         if( fail++ > 10 ) return;

         tty->print("We have it as: ");

         us->dump();

         tty->print("Verify thinks: ");

         them->dump();

         tty->cr();

       }

     }

   }

   // Check for immediate dominators being equal

   if( i >= _idom_size ) {

     if( !n->is_CFG() ) return;

     tty->print("CFG Node with no idom: ");

     n->dump();

     return;

   }

   if( !n->is_CFG() ) return;

   if( n == C->root() ) return; // No IDOM here

   assert(n->_idx == i, "sanity");

   Node *id = idom_no_update(n);

   if( id != loop_verify->idom_no_update(n) ) {

     tty->print("Unequals idoms for: ");

     n->dump();

     if( fail++ > 10 ) return;

     tty->print("We have it as: ");

     id->dump();

     tty->print("Verify thinks: ");

     loop_verify->idom_no_update(n)->dump();

     tty->cr();

   }

 }

 //------------------------------verify_tree------------------------------------

 // Verify that tree structures match.  Because the CFG can change, siblings

 // within the loop tree can be reordered.  We attempt to deal with that by

 // reordering the verify's loop tree if possible.

 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {

   assert( _parent == parent, "Badly formed loop tree" );

   // Siblings not in same order?  Attempt to re-order.

   if( _head != loop->_head ) {

     // Find _next pointer to update

     IdealLoopTree **pp = &loop->_parent->_child;

     while( *pp != loop )

       pp = &((*pp)->_next);

     // Find proper sibling to be next

     IdealLoopTree **nn = &loop->_next;

     while( (*nn) && (*nn)->_head != _head )

       nn = &((*nn)->_next);

     // Check for no match.

     if( !(*nn) ) {

       // Annoyingly, irreducible loops can pick different headers

       // after a major_progress operation, so the rest of the loop

       // tree cannot be matched.

       if (_irreducible && Compile::current()->major_progress())  return;

       assert( 0, "failed to match loop tree" );

     }

     // Move (*nn) to (*pp)

     IdealLoopTree *hit = *nn;

     *nn = hit->_next;

     hit->_next = loop;

     *pp = loop;

     loop = hit;

     // Now try again to verify

   }

   assert( _head  == loop->_head , "mismatched loop head" );

   Node *tail = _tail;           // Inline a non-updating version of

   while( !tail->in(0) )         // the 'tail()' call.

     tail = tail->in(1);

   assert( tail == loop->_tail, "mismatched loop tail" );

   // Counted loops that are guarded should be able to find their guards

   if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {

     CountedLoopNode *cl = _head->as_CountedLoop();

     Node *init = cl->init_trip();

     Node *ctrl = cl->in(LoopNode::EntryControl);

     assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );

     Node *iff  = ctrl->in(0);

     assert( iff->Opcode() == Op_If, "" );

     Node *bol  = iff->in(1);

     assert( bol->Opcode() == Op_Bool, "" );

     Node *cmp  = bol->in(1);

     assert( cmp->Opcode() == Op_CmpI, "" );

     Node *add  = cmp->in(1);

     Node *opaq;

     if( add->Opcode() == Op_Opaque1 ) {

       opaq = add;

     } else {

       assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );

       assert( add == init, "" );

       opaq = cmp->in(2);

     }

     assert( opaq->Opcode() == Op_Opaque1, "" );

   }

   if (_child != NULL)  _child->verify_tree(loop->_child, this);

   if (_next  != NULL)  _next ->verify_tree(loop->_next,  parent);

   // Innermost loops need to verify loop bodies,

   // but only if no 'major_progress'

   int fail = 0;

   if (!Compile::current()->major_progress() && _child == NULL) {

     for( uint i = 0; i < _body.size(); i++ ) {

       Node *n = _body.at(i);

       if (n->outcnt() == 0)  continue; // Ignore dead

       uint j;

       for( j = 0; j < loop->_body.size(); j++ )

         if( loop->_body.at(j) == n )

           break;

       if( j == loop->_body.size() ) { // Not found in loop body

         // Last ditch effort to avoid assertion: Its possible that we

         // have some users (so outcnt not zero) but are still dead.

         // Try to find from root.

         if (Compile::current()->root()->find(n->_idx)) {

           fail++;

           tty->print("We have that verify does not: ");

           n->dump();

         }

       }

     }

     for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {

       Node *n = loop->_body.at(i2);

       if (n->outcnt() == 0)  continue; // Ignore dead

       uint j;

       for( j = 0; j < _body.size(); j++ )

         if( _body.at(j) == n )

           break;

       if( j == _body.size() ) { // Not found in loop body

         // Last ditch effort to avoid assertion: Its possible that we

         // have some users (so outcnt not zero) but are still dead.

         // Try to find from root.

         if (Compile::current()->root()->find(n->_idx)) {